Aero-engine fault identification method and device based on cross-dimensional hybrid diagnosis

Abstract

The invention relates to an aero-engine fault identification method and device based on cross-dimensional hybrid diagnosis. The method comprises the steps that a target aero-engine data set of a target aero-engine is acquired, the target aero-engine data set comprises aero-engine data, and the aero-engine data comprises zero-dimensional data and high-dimensional data; extracting time domain features and / or frequency domain features of the 0-dimensional data as 0-dimensional feature quantities, and extracting data capable of representing high-dimensional spatial-temporal features from the high-dimensional data as high-dimensional feature quantities; performing fusion processing on the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fused feature quantity; training a fault recognition model based on the fused feature quantity; and a real-time aero-engine data set of the target aero-engine is acquired, and the actual fault type of the target aero-engine and the occurrence probability of the actual fault type are identified and obtained based on the fault identification model. By adopting the method, the fault identification precision can be improved.

Description

Technical Field

[0001] This application relates to the field of jet propulsion devices, and particularly to a method and device for fault identification of aero-engines based on cross-dimensional hybrid diagnosis. Background Art

[0002] As an important component of an aircraft, the performance of an aero-engine directly affects the operation of the aircraft. Therefore, identifying possible faults in the aero-engine and giving early warnings of the faults in a timely manner are crucial for the normal use of the aircraft.

[0003] In the related art, the fault type of an aero-engine is usually judged by whether key indicators are abnormal. However, since the abnormality of the same indicator in an aero-engine may be caused by multiple factors, judging the fault type in this way is not accurate, and the recognition accuracy of the fault type is low. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a method and device for fault identification of an aero-engine based on cross-dimensional hybrid diagnosis.

[0005] In a first aspect, this application provides a method for constructing a fault warning threshold of an aero-engine. The method includes:

[0006] Obtain a target aero-engine dataset of a target aero-engine, where the target aero-engine dataset includes various aero-engine data obtained for the target aero-engine. The aero-engine data includes 0-dimensional data and high-dimensional data, where the 0-dimensional data is obtained through an aero-engine zero-dimensional design tool, and the high-dimensional data is obtained through an aero-engine high-dimensional simulation tool. The 0-dimensional data and the high-dimensional data are data calibrated by test data;

[0007] Extract the time-domain feature and / or frequency-domain feature of the 0-dimensional data as the 0-dimensional feature quantity of the 0-dimensional data, and extract the data that can characterize the high-dimensional spatio-temporal feature from the high-dimensional data as the high-dimensional feature quantity of the high-dimensional data;

[0008] Perform a fusion process on the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fusion feature quantity;

[0009] Establish a fault identification model based on the fusion feature quantity and the fault type, and perform a training process on the fault identification model;

[0010] Obtain a real-time aero-engine dataset of the target aero-engine, and input the real-time aero-engine dataset into the trained fault identification model to identify the actual fault type of the target aero-engine and the occurrence probability of the actual fault type.

[0011] Second aspect, the present application also provides an aero-engine fault identification device based on cross-dimensional hybrid diagnosis. The device includes:

[0012] An acquisition module, configured to acquire a target aero-engine dataset of a target aero-engine, where the target aero-engine dataset includes various aero-engine data acquired for the target aero-engine, and the aero-engine data includes 0D data and high-dimensional data. The 0D data is obtained through an aero-engine zero-dimensional design tool, and the high-dimensional data is obtained through an aero-engine high-dimensional simulation tool. The 0D data and the high-dimensional data are data calibrated by test data;

[0013] An extraction module, configured to extract the time-domain feature and / or frequency-domain feature of the 0D data as the 0D feature quantity of the 0D data, and extract the data capable of characterizing the high-dimensional spatio-temporal feature from the high-dimensional data as the high-dimensional feature quantity of the high-dimensional data;

[0014] A fusion module, configured to perform a fusion process on the 0D feature quantity and the high-dimensional feature quantity to obtain a fusion feature quantity;

[0015] A training module, configured to establish a fault identification model based on the fusion feature quantity and the fault type, and perform a training process on the fault identification model;

[0016] An identification module, configured to acquire a real-time aero-engine dataset of the target aero-engine, and input the real-time aero-engine dataset into the trained fault identification model to identify the actual fault type of the target aero-engine and the occurrence probability of the actual fault type.

[0017] Third aspect, the present application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the above-mentioned method is implemented.

[0018] Fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method is implemented.

[0019] Fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the above-mentioned method is implemented.

[0020] The above-mentioned aero-engine fault identification method and device based on cross-dimensional hybrid diagnosis collect a target aero-engine data set composed of aero-engine data in multiple dimensions, fuse each aero-engine data to obtain a fused feature quantity, and after training a fault identification model based on the fused feature quantity, in application, the actually generated real-time aero-engine data can be input into the fault identification model to obtain the actual fault type and its occurrence probability of the target aero-engine, which can improve the fault judgment accuracy of the aero-engine in actual application. Description of the Drawings

[0021] Figure 1 It is a schematic flow chart of the aero-engine fault identification method based on cross-dimensional hybrid diagnosis in one embodiment;

[0022] Figure 2 It is a schematic flow chart of the aero-engine fault identification method based on cross-dimensional hybrid diagnosis in another embodiment;

[0023] Figure 3 It is a structural block diagram of the aero-engine fault identification device based on cross-dimensional hybrid diagnosis in one embodiment;

[0024] Figure 4 It is an internal structure diagram of a computer device in one embodiment. Detailed Embodiments

[0025] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0026] In one embodiment, as Figure 1 shown, a aero-engine fault identification method based on cross-dimensional hybrid diagnosis is provided. In this embodiment, the application of this method to a server is taken as an example for illustration. It can be understood that this method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is realized through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0027] Step 102, obtain a target aero-engine data set of a target aero-engine, where the target aero-engine data set includes a variety of aero-engine data obtained for the target aero-engine, and the aero-engine data includes 0-dimensional data and high-dimensional data, where the 0-dimensional data is obtained through an aero-engine zero-dimensional design tool, the high-dimensional data is obtained through an aero-engine high-dimensional simulation tool, and the 0-dimensional data and the high-dimensional data are data calibrated through test data.

[0028] In the embodiments of the present application, during the operation of the target aeroengine, the target aeroengine dataset can be periodically collected, and the target aeroengine dataset of the target aeroengine can also be obtained through a simulation method. The target aeroengine dataset can include various aeroengine data, such as 0D scalar data (temperature, pressure, rotational speed, etc.) and high-dimensional data (velocity field, temperature field), etc.

[0029] Exemplarily, when obtaining the target aeroengine dataset through a simulation method, 0D data can be obtained through aeroengine 0D design tools including but not limited to self-written programs, PROOSIS, Simcenter Amesim and other programs, and high-dimensional data can be obtained through aeroengine high-dimensional simulation tools including but not limited to self-written programs, ANSYS Fluent, FLOW-3D, CFDPro, etc. It should be noted that the aeroengine 0D design tools such as PROOSIS and Simcenter Amesim, and the aeroengine high-dimensional simulation tools such as ANSYS Fluent, FLOW-3D, and CFDPro are only used as an example, and the embodiments of the present application do not make any limitations thereto, nor does it mean that the above-mentioned aeroengine 0D design tools such as PROOSIS and Simcenter Amesim, and the aeroengine high-dimensional simulation tools such as ANSYS Fluent, FLOW-3D, and CFDPro will be used in the specific implementation process.

[0030] The data calibrated by experimental data refers to the data output based on the calibrated aeroengine 0D design tool or aeroengine high-dimensional simulation tool after calibrating the aforementioned aeroengine 0D design tool or aeroengine high-dimensional simulation tool with experimental data. The sources of experimental data can include: obtaining temperature data in 0D data by measuring the temperature of the combustion chamber and turbine through a temperature sensor (such as a thermocouple), obtaining pressure data in 0D data by acquiring the intake or exhaust pressure through a piezoelectric pressure sensor, obtaining rotational speed data in 0D data by recording the rotational speed of the rotor through a Hall effect rotational speed sensor; capturing the velocity field in the aeroengine in high-dimensional data through a particle image velocimeter or a laser Doppler velocimeter; measuring the temperature field in high-dimensional data by laser-induced fluorescence; and obtaining the surface temperature gradient of the aeroengine in high-dimensional data in real time by infrared thermal imaging.

[0031] Step 104, extract the time-domain feature and / or frequency-domain feature of the 0D data as the 0D feature quantity of the 0D data, and extract the data capable of characterizing the high-dimensional spatio-temporal feature from the high-dimensional data as the high-dimensional feature quantity of the high-dimensional data.

[0032] In the embodiments of the present application, high-dimensional feature quantities of each high-dimensional aero-engine data (high-dimensional data) can be extracted respectively by a high-dimensional feature quantity extraction model. The high-dimensional feature quantity extraction model can be composed of multiple convolutional layers, pooling layers and fully connected layers. Among them, the convolutional layer is used to extract local spatio-temporal features of the high-dimensional data, and the pooling layer is used to enhance the non-linear expression of the features extracted by the convolutional layer. An activation function (such as the ReLU activation function) can also be set after the pooling layer to activate the features output by the pooling layer, and then the activated features are input into the fully connected layer to obtain high-dimensional spatio-temporal features, and the high-dimensional spatio-temporal features are used as the high-dimensional feature quantities of the high-dimensional data.

[0033] Time-domain features or frequency-domain features of the 0-dimensional aero-engine data (0-dimensional data) can also be extracted as 0-dimensional feature quantities. The time-domain features can include, but are not limited to, the mean value, extreme value, variance, etc. of the 0-dimensional data collected at multiple moments, and the frequency-domain features can include, but are not limited to, the frequency distribution obtained after converting the 0-dimensional data to the frequency domain.

[0034] Step 106, perform a fusion process on the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fusion feature quantity.

[0035] In the embodiments of the present application, after performing a fusion process on each high-dimensional feature quantity to obtain an initial fusion feature quantity, the 0-dimensional feature quantity of each 0-dimensional data and the initial fusion feature quantity can be concatenated to obtain a fusion feature quantity. The concatenation method can be front-back concatenation, or a fully connected layer for fusing the time-domain features and frequency-domain features of each 0-dimensional data and the initial fusion feature quantity can be added after the output layer of the high-dimensional feature quantity extraction model to obtain a fusion feature quantity.

[0036] Step 108, establish a fault recognition model based on the fusion feature quantity and the fault type, and perform a training process on the fault recognition model.

[0037] In the embodiments of the present application, the fault type refers to a preset fault type that the target aero-engine may have. The fault recognition model can be any classification model, such as a support vector machine, a random forest, a deep neural network, etc. Further, the fault recognition model can also be obtained by connecting the aforementioned high-dimensional feature quantity extraction model and the classification model in series, so that when applied, the 0-dimensional data and the high-dimensional data are input into the fault recognition model to obtain the fault type.

[0038] Inputting the fusion feature quantity into the fault recognition model, the predicted fault type output by the fault recognition model can be obtained. Training the model based on the predicted fault type and the actual fault type corresponding to the fusion feature quantity can obtain a trained fault recognition model.

[0039] Step 110: Obtain the real-time aero-engine dataset of the target aero-engine, and input the real-time aero-engine dataset into the trained fault identification model to identify the actual fault type of the target aero-engine and the occurrence probability of the actual fault type.

[0040] In the embodiment of the present application, after the fault identification model is trained, the real-time aero-engine dataset of the target aero-engine can be collected during the application process. Input each aero-engine data in the real-time aero-engine dataset into the trained fault identification model (or after fusing to obtain the fused feature quantity, input the fused feature quantity into the trained fault identification model). The trained fault identification model can output the actual fault type and the occurrence probability of the actual fault type. The occurrence probability refers to the confidence level that the fault identification model believes the real-time aero-engine dataset belongs to the actual fault type.

[0041] The aero-engine fault identification method based on cross-dimensional hybrid diagnosis provided by the embodiment of the present application collects the target aero-engine dataset composed of aero-engine data of multiple dimensions, fuses each aero-engine data to obtain the fused feature quantity, and after training the fault identification model based on the fused feature quantity, in the application, the actually generated real-time aero-engine data can be input into the fault identification model to obtain the actual fault type of the target aero-engine and its occurrence probability, which can improve the fault judgment accuracy of the aero-engine in the actual application.

[0042] In one embodiment, the fusion process of the time-domain feature and / or frequency-domain feature and the high-dimensional feature quantity to obtain the fused feature quantity includes:

[0043] Determine the weights of the 0-dimensional feature quantity and the high-dimensional feature quantity respectively;

[0044] Based on the weights, splice the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain the spliced feature, and input the spliced feature into the stacked neural network. Process the features of different dimensions in the spliced feature through different branches of the stacked neural network, and fuse the features output by each branch through the fully connected layer of the stacked neural network to obtain the fused feature quantity.

[0045] In the embodiment of the present application, a stacked neural network can be designed to fuse the 0-dimensional feature quantity and the high-dimensional feature quantity. Before inputting the 0-dimensional feature quantity and the high-dimensional feature quantity into the stacked neural network, first determine the weights of each 0-dimensional feature quantity and high-dimensional feature quantity, and then splice the 0-dimensional feature quantity and the high-dimensional feature quantity based on the weights to obtain the spliced feature, which is also the input of the stacked neural network. The embodiment of the present application does not limit the splicing method, and any splicing method is applicable to the embodiment of the present application. The weights can be set manually or determined based on any adaptive weight allocation method.

[0046] The stacked neural network includes multiple branches, each of which can be used to process the features of a certain dimension in the spliced features and output the features obtained after further processing of the features. A fully connected layer is set after each branch to fuse the features output by each branch, and then a fused feature quantity is obtained.

[0047] In one embodiment, extracting the high-dimensional feature quantity of high-dimensional data includes:

[0048] Embedding the 0-dimensional data into each grid node of the grid data of the target aeroengine to obtain the basic aeroengine data;

[0049] Based on the basic aeroengine data and the values of the high-dimensional data at each grid node of the grid data of the target aeroengine, determining the target high-dimensional data corresponding to the high-dimensional data;

[0050] Extracting the high-dimensional feature quantity based on the target high-dimensional data.

[0051] In the embodiments of the present application, the 0-dimensional data can be embedded into each grid node of the target aeroengine as the fusion basis of the high-dimensional data. Embedding means copying the 0-dimensional data to each grid node. The grid data refers to the two-dimensional or three-dimensional grid data formed after grid division of the inside of the aeroengine. For the convenience of processing by the fault identification model, the grid data can be structured grid data.

[0052] After embedding the 0-dimensional data into the grid nodes, the basic aeroengine data is obtained. Then, for each high-dimensional data, the high-dimensional data can be aligned with the grid nodes, and the basic aeroengine data and the aligned high-dimensional data are fused to obtain the target high-dimensional data corresponding to this high-dimensional data.

[0053] The method for determining the value of the high-dimensional data at the grid node can be to interpolate the high-dimensional data. The interpolation methods can include cubic spline interpolation, bilinear interpolation, etc. If the dimension of the grid data is lower than the dimension of the high-dimensional data, the high-dimensional data can also be dimensionally reduced to the dimension of the grid data, and then the interpolated data after dimensional reduction is used to obtain the value of the high-dimensional data at the grid node.

[0054] Illustrated with actual examples. For example, when the grid data is two-dimensional grid data on the xz plane in a three-dimensional coordinate system and the high-dimensional data is three-dimensional velocity field data including velocity vectors in the x, y, and z directions, taking a grid node with coordinates (x’, z’) as an example, if the high-dimensional data is not reduced in dimension, the alignment method can be to interpolate the velocity field data to obtain each velocity vector corresponding to the x-axis coordinate x’ and the z-axis coordinate z’: for example, it may include (x1, y1, z1), (x2, y2, z2) …… (xn, yn, zn), etc. These velocity field data can all be used as the values of the high-dimensional data at the grid node, that is, the value is a data sequence including (x1, y1, z1), (x2, y2, z2) …… (xn, yn, zn).

[0055] If the high-dimensional data is reduced in dimension, the alignment method is to first reduce the three-dimensional velocity field data to two dimensions to obtain the reduced-dimensional velocity field data on the xz plane. The embodiments of the present application do not limit the reduction method, and any reduction method is applicable to the embodiments of the present application. Interpolating the reduced-dimensional velocity field data can obtain the two-dimensional velocity vector (such as (x1, z1)) at the coordinates (x’, z’), and this velocity vector is the value of the high-dimensional data at the grid node.

[0056] If the dimension of the high-dimensional data is less than the dimension of the grid data, the alignment method can be to align the high-dimensional data with the grid node in the dimension of the high-dimensional data through interpolation or other methods. For example, when the grid data is three-dimensional grid data and the high-dimensional data is two-dimensional temperature field data on the xz plane, for a grid node with coordinates (x’, y’, z’), the value of the high-dimensional data at the coordinates (x’, z’) on the xz plane can be determined, and this value is used as the value of the high-dimensional data at the grid node.

[0057] After obtaining the value of the high-dimensional data at the grid node, the value of the high-dimensional data at the grid node and the basic aero-engine data can be further spliced to obtain the target high-dimensional data corresponding to the high-dimensional data. Then the target high-dimensional data can be input into the high-dimensional feature quantity extraction model to obtain the high-dimensional feature quantity.

[0058] In one embodiment, a fault identification model is established based on the fusion feature quantity and the fault type, including:

[0059] Determine the labeled fault type corresponding to the fusion feature quantity;

[0060] A mapping modeling method is used to establish a mapping relationship between the fused feature quantities and the labeled fault types, forming an initial fault identification model. Among them, the mapping modeling method includes a mapping modeling method based on machine learning and a mapping modeling method based on deep learning. The mapping modeling method based on machine learning includes any one of support vector machine, random forest, and neural network. The mapping modeling method based on deep learning includes any one of convolutional neural network and recurrent neural network.

[0061] Input the fused feature quantities into the initial fault identification model to obtain the predicted fault types output by the initial fault identification model.

[0062] Train the initial fault identification model based on the difference between the labeled fault types and the predicted fault types to obtain a fault identification model.

[0063] In the embodiments of the present application, the fused feature quantities have corresponding labeled fault types. An initial fault identification model can be obtained by establishing a mapping relationship between the fused feature quantities and the labeled fault types. The establishment method of this mapping relationship includes, but is not limited to, establishing a support vector machine, random forest, neural network, convolutional neural network, recurrent neural network, etc. that can predict the fault types of the fused feature quantities based on the fused feature quantities.

[0064] After inputting the fused feature quantities into the initial fault identification model, the fault types to which the fused feature quantities predicted by the model belong in each preset fault type can be obtained. Based on whether the labeled fault type corresponding to the fused feature quantity matches the fault type predicted by the model, the loss value of the model can be calculated. Training the initial fault identification model based on the loss value can obtain a trained fault identification model. During the training process, the hyperparameters used in training the initial fault identification model can also be adjusted by means of k-fold cross-validation, etc., and the model can be prevented from overfitting by means of L2 regularization or Dropout (setting the output of random neurons to 0 during the training process).

[0065] It should be noted that multiple initial fault identification models can also be trained, and the recognition accuracy rate of each initial fault identification model can be evaluated during the training process. Then, the Stacking ensemble learning method, etc. can be used to comprehensively obtain the finally output fault type based on the recognition accuracy rate and the outputs of each initial fault identification model. The recognition accuracy rate can be evaluated by means of F1 score (F1 value), confusion matrix, etc.

[0066] In one embodiment, the above method further includes:

[0067] Determine extended fused feature quantities based on the fused feature quantities;

[0068] Identify the preset fault types corresponding to each extended fusion feature quantity based on the fault identification model, and construct the abnormal threshold of the aero-engine data corresponding to each preset fault type based on the target aero-engine data sets corresponding to the extended fusion feature quantities belonging to each preset fault type respectively.

[0069] In the embodiments of the present application, the extended fusion feature quantity may be data obtained by further experiments or simulations on the target aero-engine, or data generated based on the fusion feature quantity, or the fusion feature quantity extracted based on the extended target aero-engine data set after generating the extended target aero-engine data set based on the target aero-engine data set corresponding to the fusion feature quantity. The extended fusion feature quantity is data without labeled fault types.

[0070] Exemplarily, after conducting experiments, simulations, and actual operation data collection on the target aero-engine, the fusion feature quantity corresponding to the target aero-engine data set with a fault type label can be used as the fusion feature quantity for model training, and the fusion feature quantity corresponding to the target aero-engine data set without a fault type label can be used as the extended fusion feature quantity, so that the training of the fault identification model can be based on the labeled data. Among them, the target aero-engine data set with a fault type label can be obtained through manual labeling, by creating a specific fault type for the aero-engine in the experiment and collecting the aero-engine data under this fault type as the target aero-engine data set, or it can also be obtained by simulating a certain fault in the simulation. The target aero-engine data set without a fault type label may be a data set generated through data augmentation or other means, or a data set for which the specific fault type cannot be determined in the experiment or simulation.

[0071] After training the fault identification model, input each extended fusion feature quantity into the fault identification model to obtain the preset fault type to which each extended fusion feature quantity belongs among each preset fault type. Furthermore, the target aero-engine data sets corresponding to the extended fusion feature quantities determined by the model to belong to each preset fault type can be obtained, and the abnormal threshold of the aero-engine data corresponding to each preset fault type can be statistically obtained based on the aero-engine data in each target aero-engine data set. The abnormal threshold of the aero-engine data can be the abnormal threshold for one or more types of aero-engine data.

[0072] Exemplarily, for a preset fault type, based on different types of aero-engine data respectively, clustering can be performed on the target aero-engine data set of the preset fault type and the target aero-engine data set of normal and fault-free conditions, and one or more types of aero-engine data that can make the difference between the clustering of the preset fault type and the fault-free clustering relatively large are selected. An aero-engine data anomaly threshold is constructed based on the difference between the clustering boundary of the preset fault type and the fault-free clustering boundary. In this way, an aero-engine data anomaly threshold that can distinguish the preset fault type from the fault-free condition is obtained.

[0073] Alternatively, clustering can also be performed on the target aero-engine data sets of different preset fault types based on different types of aero-engine data respectively. Specifically, for any preset fault type, clustering processing can be performed on the target aero-engine data set corresponding to the extended fusion feature quantity belonging to the preset fault type to obtain multiple clustering clusters; an aero-engine data anomaly threshold corresponding to each preset fault type is constructed based on the clustering boundary of each clustering cluster. For example, the clustering boundary is used as the aero-engine data anomaly threshold, or the aero-engine data anomaly threshold is obtained by multiplying the clustering boundary by a certain coefficient, and so on.

[0074] Or, in another embodiment, at least one target aero-engine data category can be determined, and clustering processing is performed on the aero-engine data belonging to the target aero-engine data category in the target aero-engine data sets corresponding to each extended fusion feature quantity to obtain multiple clustering clusters; for any first preset fault type in the preset fault types, when the distance between the clustering cluster corresponding to the first preset fault type and all the clustering clusters corresponding to each second preset fault type is greater than a preset distance threshold, an aero-engine data anomaly threshold of the first preset fault type for each target aero-engine data category is constructed; where the second preset fault type is the fault type other than the first preset fault type in the preset fault types; the target aero-engine data category is re-determined, and the process jumps to the step of performing clustering processing on the aero-engine data belonging to the target aero-engine data category in the target aero-engine data sets corresponding to each extended fusion feature quantity until each preset fault type has a corresponding aero-engine data anomaly threshold.

[0075] In the embodiment of the present application, for a certain preset fault type, a target aero-engine data category that can make the clustering of the preset fault type farthest from the clustering of other preset fault types can be determined, and an aero-engine data anomaly threshold corresponding to the target aero-engine data category of the preset fault type is constructed (such as: using the value of the clustering boundary of the clustering cluster of the preset fault type as the aero-engine data anomaly threshold, or obtaining the anomaly threshold by multiplying the clustering boundary by a certain coefficient, and so on). In this way, an aero-engine data anomaly threshold that can distinguish different fault types is obtained.

[0076] The target aero-engine data category refers to the types of aero-engine data in the foregoing embodiments. For example, temperature is a target aero-engine data category, rotational speed is also a target aero-engine data category, and temperature field is also a target aero-engine data category.

[0077] Each time, one or more target aero-engine data categories can be randomly selected as the clustering basis, and in the target aero-engine data set corresponding to each extended fusion feature quantity, the aero-engine data belonging to the target aero-engine data category is obtained. Based on the aero-engine data corresponding to each extended fusion feature quantity in the target aero-engine data category, clustering processing is performed on each extended fusion feature quantity, and multiple clustering clusters can be obtained. For each preset fault type, the clustering clusters to which each extended fusion feature quantity belonging to the preset fault type belongs are determined. If there are multiple such clustering clusters, then determine the proportion of the number of extended fusion feature quantities included in each clustering cluster in all extended fusion feature quantities, and use the clustering cluster with the highest proportion as the clustering cluster corresponding to the preset fault type.

[0078] Then determine whether there is a distance greater than the preset clustering threshold between the clustering cluster corresponding to a certain preset fault type and the clustering clusters of other preset fault types. If so, it indicates that the preset fault type and other preset fault types can be better distinguished based on the target aero-engine data category. At this time, an abnormal threshold of the aero-engine data corresponding to the preset fault type in the target aero-engine data category can be constructed.

[0079] Repeat the above process multiple times until each preset fault type has a corresponding abnormal threshold of aero-engine data, and the abnormal thresholds of aero-engine data corresponding to each preset fault type can be obtained.

[0080] In one embodiment, the above method further includes:

[0081] Match each aero-engine data in the real-time aero-engine data set with the abnormal thresholds of aero-engine data corresponding to each preset fault type respectively. In the case where there is a target preset fault type with a matching abnormal threshold of aero-engine data and aero-engine data, give an early warning based on the target preset fault type.

[0082] In the embodiments of the present application, before inputting the real-time aero-engine data set into the fault recognition model in the application, an abnormal threshold judgment is first performed to determine whether to give an early warning.

[0083] Each preset fault type can correspond to one or more abnormal thresholds of aero-engine data, and each abnormal threshold of aero-engine data corresponds to one or more types of aero-engine data.

[0084] For example, for a preset fault type A, this fault type can correspond to abnormal threshold values A1 and A2 of aero-engine data. Among them, A1 includes the threshold value for the temperature of the aero-engine, the threshold value for the rotational speed of the aero-engine, and the threshold value for the pressure of the aero-engine; A2 includes the threshold value for the rotational speed of the aero-engine (this threshold value may be the same as or different from the threshold value in A1), the threshold value for the speed field of the aero-engine, and the threshold value for the temperature field of the aero-engine. If in the target aero-engine dataset, the temperature is higher than the threshold value of the aero-engine temperature in A1, the rotational speed is higher than the threshold value of the aero-engine rotational speed in A1, and the pressure is higher than the threshold value of the aero-engine pressure in A1, then the aero-engine data in the target aero-engine dataset matches A1; if in the target aero-engine dataset, the rotational speed is higher than the threshold value of the aero-engine rotational speed in A2, the speed vectors in at least a preset proportion of regions in the speed field are higher than the threshold value of the aero-engine speed field in A2, and the temperatures in at least a preset proportion of regions in the temperature field are higher than the threshold value of the aero-engine temperature field in A2, then the aero-engine data in the target aero-engine dataset matches A2. When the target aero-engine dataset matches any one of the abnormal threshold values of the aero-engine data for the preset fault type, it can be considered that the target aero-engine may have a fault of the preset fault type.

[0085] If there is a target preset fault type for which the abnormal threshold value of the aero-engine data matches the aero-engine data, then give an early warning based on the target preset fault type. In the case where there are multiple target preset fault types, the target preset fault type with the highest probability can be selected for early warning, or an early warning can be given based on all the target preset fault types. Here, the highest probability can mean that among all the abnormal threshold values of the aero-engine data for this target preset fault type, the proportion of the abnormal threshold values of the aero-engine data that match the target aero-engine dataset is the highest among all the target preset fault types; or it can also mean that the proportion of the aero-engine data in the target aero-engine dataset that exceeds the abnormal threshold values of the aero-engine data for this target preset fault type is the largest among all the target preset fault types.

[0086] If there is no target preset fault type, then no early warning may be given.

[0087] In one embodiment, the above method further includes:

[0088] Based on the fault types identified by the fault identification model, statistically calculate the occurrence probability of the fault types within any time period;

[0089] Based on the occurrence probability and the statistical confidence interval method, determine the alarm threshold for the fault types.

[0090] In one embodiment, the above method further includes: for any real-time aero-engine dataset, while the fault identification model identifies the actual fault type, automatically determine whether to give an early warning according to the alarm threshold.

[0091] In the embodiments of the present application, the alarm threshold is different from the aforementioned anomaly threshold. The aforementioned anomaly threshold refers to a numerical threshold, and a warning is given when the aero-engine data matches the anomaly threshold. The alarm threshold here refers to a probability threshold, and a warning is given when the occurrence probability of the actual fault type matches the alarm threshold. The warning method in the embodiments of the present application and the warning method in the aforementioned embodiments may be the same or different. For example, the warning method when the aero-engine data matches the anomaly threshold may be to give a pop-up prompt on the display device of the aircraft, and the warning method when the occurrence probability of the actual fault type matches the alarm threshold may be to give a sound prompt in the aircraft, etc.

[0092] During the operation of the target aero-engine, a real-time aero-engine data set can be collected, and the confidence levels of the real-time aero-engine data set judged by the fault identification model belonging to each possible fault type can be collected. The confidence level of the real-time aero-engine data set belonging to a certain fault type represents the occurrence probability of this fault type. By statistically analyzing these probabilities, the normal probability value of each fault type can be obtained. The normal probability value represents the occurrence probability that a certain fault type should have when the actual fault type is not that fault type.

[0093] Furthermore, the alarm threshold can be obtained based on the statistical confidence interval method. For example, it can be assumed that the confidence levels output by the fault identification model for each fault type follow a normal distribution, and the normal probability value is the mean in the normal distribution. Then, the 3σ principle in the normal distribution can be used to set the alarm threshold. The 3σ principle means that in the normal distribution, the probability that the data falls within the mean ± 3 times the standard deviation value is 99.73%. Referring to this principle, the sum of the normal probability value and 3 times the standard deviation value (where the standard deviation value can be selected as the default value in the normal distribution) can be used as the alarm threshold. In practical applications, if the occurrence probability of the actual fault type is greater than the alarm threshold corresponding to the actual fault type, it can be considered that the occurrence probability of the actual fault type is relatively large. Therefore, a warning can be given based on the actual fault type.

[0094] In one example, the occurrence probability can also be obtained statistically. Taking the fault types including the first fault type, the second fault type, and the third fault type as an example, assume that within a time period A, the first fault type is identified n1 times, the second fault type is identified n2 times, and the third fault type is identified n3 times. Then the occurrence probability of the first fault type within the time period A is: n1 / (n1 + n2 + n3). Similarly, the occurrence probability of the second fault type within the time period A is: n2 / (n1 + n2 + n3), and the occurrence probability of the third fault type within the time period A is: n3 / (n1 + n2 + n3). The statistical confidence interval method is a statistical method used to determine the confidence interval, and in the embodiments of the present application, no specific limitation is imposed on this statistical confidence interval method. When the occurrence probability corresponding to each fault type is determined, the occurrence probabilities corresponding to each fault type can be substituted into the statistical confidence interval method to obtain the confidence interval corresponding to the fault type (for example: substituting the occurrence probabilities into the calculation formulas for the upper limit value and the lower limit value of the confidence interval respectively), and based on the upper limit or the lower limit of the confidence interval, the alarm threshold corresponding to the fault type is determined. For example: taking the upper limit of the confidence interval as the alarm threshold for the fault type. While the fault identification model identifies the actual fault type, it will output the confidence level (i.e., the occurrence probability) for this actual fault type. Furthermore, the confidence level can be compared with the alarm threshold for this actual fault type. When the confidence level is greater than or equal to the alarm threshold, it is determined to give an early warning.

[0095] In one embodiment, as Figure 2 shown, a method for identifying faults in an aero-engine based on cross-dimensional hybrid diagnosis is provided, including:

[0096] S1, using sensors to collect 0D scalar data such as temperature, pressure, and rotational speed during the operation of the aero-engine, and using optical diagnostic equipment or computational fluid dynamics simulation to obtain high-dimensional data such as the internal three-dimensional flow field and temperature field of the aero-engine. Among them, a temperature sensor (such as a thermocouple) is used to measure the temperature of the combustion chamber and the turbine; a piezoelectric pressure sensor is used to obtain the intake / exhaust pressure; a Hall effect rotational speed sensor records the rotational speed of the rotor. A particle image velocimeter or a laser Doppler velocimeter is used to capture the flow velocity field; laser-induced fluorescence is used to measure the temperature field; infrared thermal imaging is used to obtain the surface temperature gradient in real time.

[0097] S2. Perform noise filtering, missing value filling, and normalization on the 0-dimensional data. For example, use wavelet transform or median filtering to eliminate noise, fill missing values through linear interpolation or K-nearest neighbor algorithm, and use Z-score normalization to eliminate dimensional differences. Perform grid re-partitioning on the high-dimensional data, that is, match the high-dimensional data collected on the unstructured grid data to each grid node on the structured grid data. Obtain the values of the high-dimensional data at each grid node on the structured grid data through interpolation. Embed the 0-dimensional data into the grid nodes.

[0098] S3. Extract statistical features from the 0-dimensional data, that is, 0-dimensional feature quantities.

[0099] S4. Use deep learning methods (such as using a convolutional neural network model) to extract high-dimensional feature quantities of the high-dimensional data. Input the high-dimensional data into the convolutional neural network model; extract local spatial patterns through multiple convolutional kernels of the convolutional neural network model; reduce the dimension through the pooling layer of the convolutional neural network model and enhance the non-linear expression through the ReLU activation function; gradually fuse the global features in the deep network of the convolutional neural network model; finally, output the spatial features, that is, high-dimensional feature quantities, through the fully connected layer of the convolutional neural network model.

[0100] S5. Integrate multi-dimensional features to construct a fusion feature quantity using weighted average, attention mechanism, or stacked feature fusion network. Specifically, splice the time-domain features and / or frequency-domain features and the high-dimensional feature quantities, and combine adaptive weighting to assign different-dimensional weights; introduce the attention mechanism to dynamically enhance important fault-related features; construct a stacked network, separate and process multi-modal data and then fuse through the cross-dimensional fully connected layer, and finally output the fusion feature quantity.

[0101] S6. Train at least one initial fault recognition model based on the fusion feature quantity to obtain a trained fault recognition model; the initial fault recognition model can include a support vector machine, a random forest, and a deep neural network. Among them, the support vector machine is used to map non-linear relationships using kernel functions, the random forest is used to perform classification through feature importance screening, and the deep neural network is used to separately process 0-dimensional data and high-dimensional data through multiple input branches and output the fault category through full connection in the fusion layer.

[0102] In S7, during the training process, fused feature quantities with fault type labels can be adopted, and the hyperparameters of the model (such as learning rate, regularization coefficient) can be optimized through k-fold cross-validation. Grid search or Bayesian optimization can be used to select the optimal parameters; data augmentation (such as adding noise, spatial transformation) can be combined to enhance the diversity of the fused feature quantities; L2 regularization or Dropout can be introduced to prevent overfitting; Stacking ensemble learning can be used to fuse the prediction results of multiple fault recognition models, and the weight of each model is determined by the recognition accuracy rate evaluated based on the F1-score and the confusion matrix, so as to improve the generalization ability and diagnostic accuracy of the fault recognition model.

[0103] In S8, the alarm threshold of key feature values (that is, certain key aero-engine data types) is statistically obtained based on the output of the fault recognition model.

[0104] In S9, in the application, it is determined whether to give an early warning by comparing with the threshold, and the fault type of the aero-engine and the confidence level of the fault type are determined through the fault recognition model. Further, the diagnostic result for the fault type can also be obtained, such as which abnormal aero-engine data causes the fault type, etc. The diagnostic result can be obtained after other models such as large language models process the output of the fault recognition model.

[0105] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments, and the execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or steps or stages in other steps.

[0106] Based on the same inventive concept, the embodiment of the present application also provides a cross-dimensional hybrid diagnosis-based aero-engine fault recognition device for implementing the above-mentioned cross-dimensional hybrid diagnosis-based aero-engine fault recognition method. The implementation solutions provided by this device to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the cross-dimensional hybrid diagnosis-based aero-engine fault recognition device provided below can refer to the limitations on the cross-dimensional hybrid diagnosis-based aero-engine fault recognition method in the above text, and will not be repeated here.

[0107] In one embodiment, as Figure 3As shown, a fault identification device 300 for an aero-engine based on cross-dimensional hybrid diagnosis is provided, including: an acquisition module 302, an extraction module 304, a fusion module 306, a training module 308, and an identification module 310, where:

[0108] The acquisition module 302 is configured to acquire a target aero-engine dataset of a target aero-engine, where the target aero-engine dataset includes various aero-engine data acquired for the target aero-engine, and the aero-engine data includes 0-dimensional data and high-dimensional data. The 0-dimensional data is obtained through an aero-engine zero-dimensional design tool, and the high-dimensional data is obtained through an aero-engine high-dimensional simulation tool. The 0-dimensional data and the high-dimensional data are calibrated data through test data;

[0109] The extraction module 304 is configured to extract the time-domain feature and / or frequency-domain feature of the 0-dimensional data as the 0-dimensional feature quantity of the 0-dimensional data, and extract data capable of characterizing high-dimensional spatio-temporal features from the high-dimensional data as the high-dimensional feature quantity of the high-dimensional data;

[0110] The fusion module 306 is configured to perform fusion processing on the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fusion feature quantity;

[0111] The training module 308 is configured to establish a fault identification model based on the fusion feature quantity and the fault type, and perform training processing on the fault identification model;

[0112] The identification module 310 is configured to acquire a real-time aero-engine dataset of the target aero-engine, and input the real-time aero-engine dataset into the trained fault identification model to identify the actual fault type of the target aero-engine and the occurrence probability of the actual fault type.

[0113] In one embodiment, the fusion module 306 is further configured to:

[0114] Determine the weights of the 0-dimensional feature quantity and the high-dimensional feature quantity respectively;

[0115] Perform splicing processing on the 0-dimensional feature quantity and the high-dimensional feature quantity based on the weights to obtain a spliced feature, input the spliced feature into a stacked neural network, process features of different dimensions in the spliced feature through different branches of the stacked neural network, and fuse the features output by each branch through the fully connected layer of the stacked neural network to obtain a fusion feature quantity.

[0116] In one embodiment, the training module 308 is further configured to:

[0117] Determine the labeled fault type corresponding to the fusion feature quantity;

[0118] A mapping modeling method is used to establish a mapping relationship between the fusion feature quantity and the labeled fault type, forming an initial fault identification model; wherein, the mapping modeling method includes a mapping modeling method based on machine learning and a mapping modeling method based on deep learning; the mapping modeling method based on machine learning includes any one of support vector machine, random forest, and neural network; the mapping modeling method based on deep learning includes any one of convolutional neural network and recurrent neural network;

[0119] The fusion feature quantity is input into the initial fault identification model, and the predicted fault type output by the initial fault identification model is obtained;

[0120] The initial fault identification model is trained based on the difference between the labeled fault type and the predicted fault type to obtain a fault identification model.

[0121] In one embodiment, the above device further includes:

[0122] A statistics module, configured to count the occurrence probability of the fault type within any time period based on the fault type identified by the fault identification model;

[0123] A determination module, which determines the alarm threshold of the fault type based on the occurrence probability and the statistical confidence interval method.

[0124] In one embodiment, for any one of the real-time aero-engine datasets, when the fault identification model identifies the actual fault type, it automatically determines whether to give an early warning according to the alarm threshold.

[0125] Each module in the above device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0126] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 4As shown. The computer device includes a processor, a memory, and a network interface connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, it implements a method for identifying faults in an aeroengine based on cross-dimensional hybrid diagnosis.

[0127] Those skilled in the art can understand that Figure 4 the structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0128] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the steps in the above method embodiments are implemented.

[0129] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0130] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0131] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been permitted by the user or fully permitted by all parties.

[0132] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., and are not limited thereto.

[0133] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0134] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. An aircraft engine fault identification method based on cross-dimensional hybrid diagnosis, characterized in that: The method comprises: Acquire a target aircraft engine data set of a target aircraft engine, the target aircraft engine data set comprising a plurality of aircraft engine data acquired for the target aircraft engine, the aircraft engine data comprising 0-dimensional data and high-dimensional data, wherein the 0-dimensional data is acquired through an aircraft engine zero-dimensional design tool, the high-dimensional data is acquired through an aircraft engine high-dimensional simulation tool, and the 0-dimensional data and the high-dimensional data are data calibrated through test data; Extracting time domain features and / or frequency domain features of the 0-dimensional data as 0-dimensional feature quantities of the 0-dimensional data, and extracting data capable of representing high-dimensional spatiotemporal features from the high-dimensional data as high-dimensional feature quantities of the high-dimensional data; Performing fusion processing on the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fused feature quantity; Establishing a fault identification model based on the fused feature quantity and the fault type, and training the fault identification model; A real-time aircraft engine data set of the target aircraft engine is obtained, and the real-time aircraft engine data set is input into the trained fault identification model to identify the actual fault type of the target aircraft engine and the probability of occurrence of the actual fault type.

2. The method according to claim 1, characterized in that The fusing the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fused feature quantity includes: Determining weights of the 0-dimensional feature quantity and the high-dimensional feature quantity respectively; The 0-dimensional feature quantity and the high-dimensional feature quantity are spliced ​​based on the weight to obtain a spliced ​​feature, and the spliced ​​feature is input into a stacked neural network. Features of different dimensions in the spliced ​​feature are processed by different branches of the stacked neural network, and the features output by each branch are fused through a fully connected layer of the stacked neural network to obtain a fused feature quantity.

3. The method according to claim 1, characterized in that The establishing of a fault identification model based on the fused feature quantity and the fault type comprises: Determine the type of labeled fault corresponding to the fused feature quantity; A mapping modeling method is used to establish a mapping relationship between the fused feature quantity and the labeled fault type to form an initial fault identification model; wherein the mapping modeling method includes a mapping modeling method based on machine learning and a mapping modeling method based on deep learning; the mapping modeling method based on machine learning includes any one of a support vector machine, a random forest, and a neural network; the mapping modeling method based on deep learning includes any one of a convolutional neural network and a recurrent neural network; Inputting the fused feature quantity into the initial fault identification model to obtain the predicted fault type output by the initial fault identification model; The initial fault identification model is trained based on the difference between the marked fault type and the predicted fault type to obtain a fault identification model.

4. The method according to claim 1, characterized in that The method further comprises: Based on the fault type identified by the fault identification model, the probability of occurrence of the fault type in any time period is counted; An alarm threshold of the fault type is determined based on the occurrence probability and statistical confidence interval method.

5. The method according to claim 4, characterized in that The method further comprises: For any of the real-time aircraft engine data sets, the fault identification model automatically determines whether to issue a warning based on the alarm threshold while identifying the actual fault type.

6. An aircraft engine fault identification device based on cross-dimensional hybrid diagnosis, characterized in that: The device comprises: an acquisition module, configured to acquire a target aeroengine data set of a target aeroengine, wherein the target aeroengine data set includes a plurality of aeroengine data acquired for the target aeroengine, wherein the aeroengine data includes 0-dimensional data and high-dimensional data, wherein the 0-dimensional data is acquired through an aeroengine zero-dimensional design tool, and the high-dimensional data is acquired through an aeroengine high-dimensional simulation tool, and the 0-dimensional data and the high-dimensional data are data calibrated through test data; An extraction module, used to extract the time domain features and / or frequency domain features of the 0-dimensional data as the 0-dimensional feature quantity of the 0-dimensional data, and to extract data capable of representing the high-dimensional spatiotemporal features from the high-dimensional data as the high-dimensional feature quantity of the high-dimensional data; A fusion module, used for fusing the 0-dimensional feature quantity and the high-dimensional feature quantity to obtain a fused feature quantity; A training module, used to establish a fault identification model based on the fused feature quantity and the fault type, and to perform training processing on the fault identification model; The identification module is used to obtain a real-time aircraft engine data set of the target aircraft engine, and input the real-time aircraft engine data set into the trained fault identification model to identify the actual fault type of the target aircraft engine and the probability of occurrence of the actual fault type.

7. The device according to claim 6, characterized in that The fusion module is also used for: Determining weights of the 0-dimensional feature quantity and the high-dimensional feature quantity respectively; The 0-dimensional feature quantity and the high-dimensional feature quantity are spliced ​​based on the weight to obtain a spliced ​​feature, and the spliced ​​feature is input into a stacked neural network. Features of different dimensions in the spliced ​​feature are processed by different branches of the stacked neural network, and the features output by each branch are fused through a fully connected layer of the stacked neural network to obtain a fused feature quantity.

8. The device according to claim 6, characterized in that The training module is also used to: Determine the type of labeled fault corresponding to the fused feature quantity; A mapping modeling method is used to establish a mapping relationship between the fused feature quantity and the labeled fault type to form an initial fault identification model; wherein the mapping modeling method includes a mapping modeling method based on machine learning and a mapping modeling method based on deep learning; the mapping modeling method based on machine learning includes any one of a support vector machine, a random forest, and a neural network; the mapping modeling method based on deep learning includes any one of a convolutional neural network and a recurrent neural network; Inputting the fused feature quantity into the initial fault identification model to obtain the predicted fault type output by the initial fault identification model; The initial fault identification model is trained based on the difference between the marked fault type and the predicted fault type to obtain a fault identification model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images